ExPO-HM: Learning to Explain-then-Detect for Hateful Meme Detection

Conference: ICLR 2026 arXiv: 2510.08630 Code: GitHub Area: Interpretability Keywords: hate speech detection, multimodal, GRPO, curriculum learning, conditional decision entropy, interpretability

TL;DR

ExPO-HM is proposed, inspired by the training pipeline of human content moderators. By combining policy manual SFT warm-up, GRPO curriculum learning, and a Conditional Decision Entropy (CDE) reward, it is the first Explain-then-Detect system to comprehensively surpass direct detection baselines across binary classification, fine-grained classification, and reasoning quality in hateful meme detection, achieving up to 15–17% F1 improvement.

Background & Motivation

Hateful meme detection is a highly challenging online content moderation task. Existing approaches fall into two main paradigms:

Direct Detection: Outputs only a binary label (hateful/benign). Representative works such as RA-HMD rely on CLIP-based methods, achieving strong performance but providing no explanations, which does not satisfy real-world moderation requirements.

Explain-then-Detect: Generates natural language explanations before classification. However, existing systems of this type (e.g., LOREHM, U-CoT+), which use CoT prompting or agent frameworks, perform worse than simple SFT baselines. Even post-training methods such as GRPO fail to close this gap.

The authors identify two key issues:

Model explanations omit critical cues: Policy-relevant information such as attack targets and attack types is not considered as plausible explanatory hypotheses by the model.

Binary reward signals are insufficient to guide reasoning: Just as human annotators cannot learn effectively from yes/no labels alone, models also require more fine-grained feedback.

The central analogy is the human moderator training pipeline—first studying a detailed policy manual, then practicing from fine-grained category judgments to binary decisions—which motivates the design of ExPO-HM.

Method

Overall Architecture

ExPO-HM consists of three stages that simulate the human moderator training process:

1. SFT-PM Warm-up: Policy manual-enhanced supervised fine-tuning, teaching the model to understand moderation policies.
2. GRPO-CL: GRPO with curriculum learning, progressing from fine-grained to binary classification.
3. CDE Reward: Conditional Decision Entropy as a proxy reward for reasoning quality.

Key Designs

SFT Policy Manual Warm-up (SFT-PM): Fine-grained dataset labels are converted into structured policy manuals used as input prompts, with each policy item accompanied by descriptions from annotation guidelines. The LMM is trained with policy manual-augmented inputs, and the target responses are fine-grained labels. Human-written gold-standard explanations are not used (off-policy usage leads to worse performance).

GRPO Curriculum Learning (GRPO-CL): A simple 50/50/50 strategy is adopted: - First 50% of training steps: fine-grained data only (encouraging reasoning exploration) - Remaining 50%: 50/50 mixture of fine-grained and binary classification data

Key effect: Standard GRPO produces an average response length of only 28 tokens for binary classification, whereas GRPO-CL nearly doubles this to 52 tokens, indicating more detailed reasoning.

Conditional Decision Entropy (CDE):

Definition: Given explanation \(\mathbf{e}\) and input \(\mathbf{x}\), CDE is the entropy of the decision conditioned on the explanation:

\[H(d \mid \mathbf{e}, \mathbf{x}) = -\mathbb{E}_{d \sim \pi_\theta(\cdot|\mathbf{e},\mathbf{x})}[\log \pi_\theta(d \mid \mathbf{e}, \mathbf{x})]\]

Rationale: Good reasoning should lead to clear and correct decisions (low entropy), while poor reasoning produces confusion (high entropy).

CDE Reward Design:

\[r_{\text{CDE}}(h, \delta) = \delta \cdot f_{\text{correct}}(h) + (1-\delta) \cdot f_{\text{wrong}}(h)\]

where \(\delta = \mathbf{1}[d = d^*]\) indicates prediction correctness. Correct and confident (low CDE) → reward; incorrect but confident → penalty (coefficient \(\rho\)); incorrect but uncertain → tolerated.

Loss & Training

Total reward: \(r = r_{\text{format}} + r_{\text{acc}} + w \cdot r_{\text{CDE}}\)

where \(r_{\text{format}} \in \{0,1\}\) checks output format and \(r_{\text{acc}} \in [0,1\}\) measures prediction correctness. Standard GRPO clipped surrogate loss and KL regularization are applied. Default hyperparameters: \(a=0.1\), \(b=0.5\), \(w=0.2\), \(\rho=0.25\).

Key Experimental Results

Main Results

Evaluation is conducted on three datasets—HatefulMemes, MAMI, and PrideMM—using Qwen2.5-VL-3B and 7B as backbone models.

Qwen2.5-VL-7B results on HatefulMemes:

Method	Binary F1	Attack F1	Target F1	LLM Judge	CDE ↓
Zero-shot	65.9	44.7	64.5	5.0	0.33
SFT	74.5	58.4	69.4	5.0	0.33
DPO	73.6	63.2	66.6	4.9	0.32
GRPO	74.5	61.2	64.5	5.2	0.26
RA-HMD (SOTA direct detection)	80.2	—	—	5.5	—
ExPO-HM	81.1	75.6	77.2	6.2	0.03

ExPO-HM is the first Explain-then-Detect system to comprehensively surpass the direct detection SOTA (RA-HMD), while substantially leading in reasoning quality.

Cross-dataset consistency (7B model):

Dataset	GRPO Binary F1	ExPO-HM Binary F1	Gain
HatefulMemes	74.5	81.1	+6.6
MAMI	76.8	82.3	+5.5
PrideMM	73.2	78.7	+5.5

Ablation Study

#	SFT-PM	GRPO-CL	CDE	Binary F1	Attack F1	Target F1	LLM ↑	CDE ↓
1	-	-	-	74.5	61.2	64.5	5.2	0.263
2	✓	-	-	75.8	70.8	70.2	5.6	0.092
3	✓	✓	-	78.4	74.3	76.1	5.8	0.056
4	✓	✓	✓	81.1	75.6	77.2	6.2	0.026

All three components contribute: SFT-PM substantially improves fine-grained metrics, GRPO-CL further advances all metrics, and CDE markedly enhances reasoning quality (LLM Judge 5.8→6.2).

Key Findings

1. Explain-then-Detect surpasses Direct Detection for the first time: All prior systems of this type fall short of the SFT baseline.
2. CDE strongly correlates with LLM-Judge: Pearson \(r=-0.78\), Spearman \(\rho=-0.81\) (\(p<0.001\)).
3. SFT warm-up strategy is critical: Binary-only SFT performs worse than the no-warm-up baseline after the RL stage.
4. CDE does not cause policy entropy collapse: Overall policy entropy is comparable to baselines without CDE.
5. Human evaluation validates the approach: ExPO-HM achieves 100% logical consistency vs. 96% for GRPO, with helpfulness scores of 2.2 vs. 1.6.

Highlights & Insights

1. The analogy to human annotator training is highly precise: The progressive pipeline—policy manual → fine-grained practice → binary judgment—translates naturally into the algorithm design.
2. CDE is an excellent proxy for reasoning quality: Conceptually simple (conditional entropy), strongly correlated with human evaluation, and usable as a differentiable reward signal.
3. Key finding: good SFT does not necessarily lead to good RL: Binary SFT achieves the best SFT-stage performance but yields the worst results after RL.
4. Consistency across three datasets: The method generalizes well across different hateful content types.
5. Exceptionally comprehensive evaluation: Ablations, warm-up strategy comparisons, CDE analysis, calibration analysis, and human evaluation are all included.

Limitations & Future Work

1. Limited dataset scale: Annotated hateful meme data, especially with explanations, is extremely scarce.
2. Single-turn interaction: Only single-turn reasoning is evaluated; multi-turn moderation dialogue scenarios are not considered.
3. Cultural dependency: Moderation policies are highly culture-specific; cross-cultural applicability is not validated.
4. Backbone model constraints: Validation is limited to Qwen2.5-VL 3B/7B.
5. The framework is extensible to other content moderation tasks (e.g., misinformation detection, cyberbullying identification).

Related Work & Insights

• RA-HMD (Mei et al., 2025): The previous SOTA direct detection method, which ExPO-HM is the first to surpass.
• LOREHM (Huang et al., 2024): A reasoning agent framework based on LLaVA-Next-34B.
• GRPO (Shao et al., 2024): ExPO-HM builds upon this by incorporating the CDE reward and curriculum learning.
• Insight: The "reasoning quality proxy" concept underlying CDE is generalizable to other tasks requiring interpretable reasoning.

Rating

• Novelty: ⭐⭐⭐⭐ — The CDE concept is original, and the curriculum learning strategy is elegantly designed.
• Technical Depth: ⭐⭐⭐⭐ — The mapping from the human training pipeline to concrete algorithmic design is complete and coherent.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ — Three datasets, multiple baselines, ablations, and human evaluation make for an exceptionally thorough evaluation.
• Writing Quality: ⭐⭐⭐⭐ — Motivation is clear and experiments are well organized.
• Value: ⭐⭐⭐⭐ — Directly applicable to content moderation in practice.
• Overall Recommendation: ⭐⭐⭐⭐ (4/5)

Related Papers

• [CVPR 2026] Language Models Can Explain Visual Features via Steering
• [ICLR 2026] Behavior Learning (BL): Learning Hierarchical Optimization Structures from Data
• [ICLR 2026] Decoupling Dynamical Richness from Representation Learning: Towards Practical Measurement
• [ICLR 2026] Decomposing Representation Space into Interpretable Subspaces with Unsupervised Learning
• [ICLR 2026] Tokenizing Single-Channel EEG with Time-Frequency Motif Learning